Jefferies x Euronext European Disruptive Tech Conference | Artificial Intelligence
This article draws on the Artificial Intelligence panel at the Jefferies x Euronext European Disruptive Tech Conference, moderated by Jimmy Williams of Jefferies, with Alex Fortescue-Webb (Global Head of Legal Engineering, Legora) and Harshit Krishna (Vice President, Multiverse Computing).
The first wave of enterprise AI was driven by curiosity. Companies experimented, ran pilots and tested what large language models could do. The industry is now shifting to a more demanding phase. Buyers now want measurable returns, finance teams are scrutinizing compute bills, and the businesses gaining ground are those that can show value at a sustainable cost.
Return on investment now leads the conversation
Eighteen months ago, a demonstration was often enough to win over a prospective customer. Legora found that many legal teams had experimented with early chatbots but had little sense of what AI could do for their business. Seeing the platform in action sparked their imagination and shortened the sales process.
That initial excitement has given way to questions about impact. Measuring return is difficult, particularly in a profession where outcomes vary from one matter to the next. Alex Fortescue-Webb (Global Head of Legal Engineering, Legora) advises clients to focus on the order of magnitude. “It doesn’t matter if it’s 700% ROI or 725% ROI,” he says. “What matters is knowing whether the gain is in the hundreds or closer to 50%.”
Efficiency has reached the boardroom
For Multiverse Computing, cost has become the defining issue for enterprise AI. Harshit Krishna (Vice President, Multiverse Computing) described a concern that entered boardrooms around six months ago and now shapes AI strategy inside the CFO’s office. As companies scale from a handful of users to thousands, the bill for running models can overwhelm the business case.
Multiverse addresses this by compressing models, typically by 50 to 60% while retaining general-purpose accuracy, and much further when a model is specialized. A legal assistant, Krishna noted, has no need to know French Caribbean history to do its job well. Smaller models can run on a company’s own servers or devices, which also matters for organizations in defense, financial services and telecoms, where sensitive data cannot leave the premises.
Open-source models are central to this shift. Multiverse estimates that around 80% of enterprises using AI now rely on at least some open-source models, and that usage has roughly quadrupled over the past year. Revenue still accrues mainly to the large closed-model providers, but a growing share of actual workloads runs elsewhere. The practical answer for many enterprises will be a hybrid, with time-critical or accuracy-critical tasks sent to frontier models and the bulk of everyday work handled locally.
Specialist applications are holding their ground
Investors often ask whether the large AI labs will simply absorb vertical markets such as law, as they have begun to do in software coding. Legora’s experience points to a slower path, at least for now. Frontier labs have launched legal offerings with considerable fanfare, but Fortescue-Webb has seen little traction among customers.
The reason lies in how lawyers work. Legal tasks vary widely, and practitioners are short on time and have little appetite for building their own or testing new tools. Adoption depends on a carefully designed interface and close support. Legora’s legal engineers, all former lawyers, work alongside client teams to reflect their particular workflows, from how they run a transaction to the deal points they track. Fortescue-Webb compares it to a heavy flywheel that is hard to start but keeps turning once it gets going.
For the frontier labs, which must keep funding enormous compute budgets, a complex and comparatively modest market may simply not be the best use of resources.
AI is also reshaping the businesses that adopt it. Law is a highly competitive services market with low switching costs. The organizational effects are arriving more slowly. Some firms are exploring a move from the traditional pyramid towards a narrower “pencil” structure with fewer junior lawyers per partner. At the same time, the volume of legal work is rising in some practice areas, and top firms can now take on mid-market mandates profitably. Whether total demand for lawyers grows or shrinks depends on which of those trends proves stronger. Human oversight will remain essential in any case, since in most jurisdictions advice without a lawyer in the loop does not count as legal advice.
Pricing models are still being worked out
AI may be the first software market where a company can grow itself into bankruptcy if it prices incorrectly. Much depends on the delivery model.
Because Multiverse’s models run on customers’ own hardware, the company does not bear the cost of compute. That allows it to charge recurring license fees per device or per model and earn very high gross margins, while customers save on compute costs. Over time, it expects to link fees more closely to usage and outcomes as its models become more widely deployed.
Legora has moved to consumption-based pricing, which suits its market well. Law firms have always billed by the hour, so paying for usage feels familiar. Fortescue-Webb also cautions against focusing too narrowly on model costs, since having a paralegal do the same work can cost 10 to 100 times more.
Adoption is the larger untapped opportunity
Asked what the market underestimates, experts point to adoption. Fortescue-Webb described an iceberg, with today’s advanced use cases only the visible tip. Krishna recalls loading from floppy disks and now carrying terabytes in a pocket. He expects AI to follow the same path, with people running several models locally on everyday devices. If only a fraction of consumers use what current models can already do, growth over the next five years can come from adoption alone, with no improvement in the models needed.
Model capability will keep improving, but the next stage of value creation depends on deployment: lowering the cost of running AI, fitting it into real workflows and pricing it in ways customers understand.
